Steps of AI adoption

AI Skill Levels

Six ways of working with AI, from a safe first task to a system that carries learning forward. Each row shows the human role, coordination model, practical constraints, useful products, and guardrails for that level.

The roles and transitions are practical examples, not requirements for earning a level.

Proxon AI Skill Levels, their operating patterns, products, and guardrails
Step & your role	Agents	What it looks like	What’s the bottleneck	Products that help	Guardrails
L0NewTypical roleExplorer	Human-guided AI use	Safe first use Knows what work is safe for AI, what data should stay out, and which starter tasks are useful. You choose one bounded, low-risk task and learn where AI is useful before expanding its reach. Unlock: A blank page or unfamiliar task becomes a safe first experiment.	Choosing a task that is useful, small, and easy to check Knowing which data, decisions, and systems should stay out	ChatGPT, Claude, Gemini, or Microsoft 365 Copilot Proxon Data Collection for choosing enabled sources and capture depth	Use approved tools and company accounts Keep secrets, regulated data, and sensitive customer data out Require acknowledgment before content-aware collection Start with reversible work and verify every result
How to move from L0 to L1: Move from trying AI once to using dialogue repeatedly for real work.
L1ChatTypical roleCollaborator	Conversational assistant	Ad-hoc assistance Uses AI as a conversational helper for summarizing, drafting, rewriting, and explaining work. You ask, refine, and judge each response while AI helps summarize, draft, rewrite, or explain. Unlock: Routine thinking and writing move faster, while you remain in the loop for every turn.	Restating the same background and constraints in each conversation Moving answers back into the document, ticket, or system where the work lives	ChatGPT Enterprise, Claude Enterprise, Gemini Enterprise, or Microsoft 365 Copilot Proxon My Usage for captured activity, tools, recurring agents, tokens, and cost estimates	Review before publishing, sending, or making a decision Use enterprise retention, access, and model settings Treat visibility as coverage-dependent on configured collection sources Set spend limits and make accountable use visible
How to move from L1 to L2: Bring the work and its context into the session instead of moving isolated answers around by hand.
L2Contextual WorkTypical roleDirector and editor	Context-aware agent	AI works inside artifacts Brings files, repos, docs, or project context into the session so AI can make bounded changes. You bring files, repositories, documents, and tools into the session, define the boundary, and review changes in place. Unlock: AI can complete a bounded piece of real work instead of returning an answer you must transfer by hand.	Supplying enough current context without exposing unrelated information Keeping changes inside the requested boundary and reviewing the diff	OpenAI Codex, Claude Code, Cursor, or GitHub Copilot MCP servers and approved connectors for governed context Proxon AI Catalog and Cost Intelligence for observed tools, data sources, activity, and available cost signals	Grant least-privilege access to files, tools, and data sources Require review for writes, commands, and external actions Treat observed catalog entries as read-only and coverage-dependent Keep credentials out of prompts, logs, and generated artifacts
How to move from L2 to L3: Decompose the work into independent streams, then compare and integrate the results.
L3OrchestrateTypical roleOrchestrator	Parallel specialist agents	Parallel and adversarial AI Splits work across researcher, drafter, critic, and reviewer roles before final human approval. You split the work into independent research, drafting, critique, and review streams, then reconcile their outputs. Unlock: Parallel effort and adversarial review increase throughput without giving up final human judgment.	Decomposing work so streams are independent and collectively complete Resolving conflicting outputs and integrating them into one decision	OpenAI Codex multi-agent work and Claude Code subagents Microsoft Copilot Studio for role-based agents and orchestration Proxon Adoption and AI Skill Levels for captured concurrency and multi-surface signals	Isolate workspaces and give every stream a clear scope Use independent tests, reviews, or evaluations before integration Keep human review in the team's own working process Escalate irreversible or high-impact actions to a named human
How to move from L3 to L4: Turn a successful recurring pattern into an owned workflow with a trigger and an exception path.
L4AutomateTypical roleWorkflow owner	Workflow-bound agents	Reusable workflows and triggers Turns repeated AI work into governed workflows with triggers, permissions, and approval gates. You turn a proven pattern into a named workflow with triggers, permissions, approvals, monitoring, and an exception path. Unlock: Repeated AI work can run consistently without rebuilding the process every time.	Making the workflow reliable across normal and exceptional cases Choosing safe triggers, owners, budgets, and approval boundaries	Microsoft Copilot Studio, n8n, Zapier, or Make GitHub Actions and enterprise schedulers for triggered work Proxon Workflows for registering named definitions with trigger labels and ordered steps Proxon AI Skill Levels for supported scheduled or triggered markers	Assign a named owner, purpose, budget, and permission set Version every workflow and provide pause and rollback controls Keep execution permissions and approvals in the automation platform Log runs, approvals, exceptions, and material outputs
How to move from L4 to L5: Let later work reuse durable context, then curate and evaluate what the system remembers.
L5LoopTypical roleSystem steward	Memory-linked agent system	Recursive improvement system Feeds outcomes, exceptions, and human feedback back into memory so every run improves the next. You connect durable memory to repeated work so later runs can reuse context, exceptions, and reviewed feedback. Unlock: The system starts each run with what prior runs learned instead of starting from zero.	Keeping memory accurate, relevant, retrievable, and free of sensitive residue Separating real improvement from accumulated noise or drift	Agent memory and governed knowledge stores Evaluation and observability tools for outcomes and regressions Proxon AI Skill Levels for memory read and write patterns from supported emitters where marker fidelity is available	Control who can read, write, retain, and delete memory Evaluate outcomes and review drift, exceptions, and stale knowledge Do not assume that observed memory use proves improvement Keep a human owner and a kill switch for the complete loop